AI Drafting for Manufacturing Maintenance Work Orders | Asset, Evidence, Safety

A manufacturer considering AI drafting for maintenance work orders should begin with a governance question: who can verify the equipment, source document, and work authorization? Maintenance teams often search manuals, job plans, asset records, and material lists before writing a work order. AI can assemble a reviewable draft. It must not turn an uncertain asset match, obsolete instruction, or missing safety plan into a confident work command. This guide explains an implementation path for factories in Thailand and elsewhere. IBM Maximo Manage documentation provides an example of CMMS data structures; it is not a product requirement. The OSHA regulation cited below is a US example, not a statement of Thai law. Local safety and legal requirements need confirmation by qualified personnel.

Define what AI may draft

A maintenance work order can contain an asset, location, reported condition, tasks, labor and craft, parts, tools, timing, safety plan, approval, and actual work records. Divide fields into three groups. Read verified IDs and codes from systems of record. Let AI propose candidate descriptions and links, with evidence. Reserve work authorization, hazardous energy isolation, on-site hazard assessment, and final approval for authorized people. IBM describes a job plan as a detailed description of work that may contain tasks, estimated labor and hours, materials, services, tools, and associated safety plans. Applying one to a work order creates a work plan that can be changed without altering the source job plan. See IBM Job Plans and Job plans and work orders.

Connect five sources before drafting

First, connect the asset register. Asset ID, physical location, line, model, serial number, and current configuration matter. A local nickname such as “pump three” may map to more than one asset. Maintain an alias table and require a human selection if the match is uncertain. Do not retrieve a machine procedure until the asset is resolved.

Second, manage approved documents with document ID, revision, effective date, and applicability. A manufacturer manual, site procedure, drawing, and change notice may disagree because the machine has been modified. Cite the exact section and a working link so an approver can examine the original. A search result alone is insufficient. Our related guide on AI equipment manual search covers finding documents; this article covers converting verified evidence into an individual work order.

Third, connect approved job plans and work history. A completed past work order is evidence, but it is not automatically an approved standard. Compare it with the current job plan. IBM explains that work orders may reference assets or locations and include job plans, safety plans, planned labor, materials, tools, and services. See Creating work orders.

Fourth, connect parts, tools, and people data. “One bearing” is not an item code. Check specifications, approved substitutes, stock location, reservations, calibration, and craft qualification where relevant. The draft must distinguish candidates from confirmed resources. AI cannot infer whether an employee holds a valid qualification or whether a part is physically available.

Fifth, connect approved safety documents and permits. Hazardous energy may be electrical, mechanical, hydraulic, pneumatic, thermal, or another form. The US OSHA 1910.147 regulation addresses control of hazardous energy for covered servicing and maintenance, including documented procedures, training, and inspections. It is a US rule, not Thai legal advice. The design lesson is to link the applicable asset-specific, approved procedure and authorized personnel. AI must not make up isolation points or release-to-work decisions. See OSHA 1910.147.

Keep an evidence chain on every field

A draft should show where each significant value came from: asset register record for an asset ID, approved job-plan revision for a task or part, document section for a technical instruction. Store document ID, revision, section, effective scope, retrieval time, and the selected value. If sources conflict, flag the conflict for a person; do not ask the model to choose the most plausible one. Preserve a snapshot of the sources seen at approval, so later revisions do not erase the decision trail.

A practical review view puts the original request beside the draft and its evidence. The reviewer should be able to open every source and see which words were quoted, summarized, or generated as a proposed explanation. NIST’s Generative AI Profile addresses confabulation and other risks of plausible but false output. Citations help review, but a citation alone does not prove that a source applies to the specific machine.

Workflow from request to CMMS

Use the sequence request → asset match → current source retrieval → structured draft → field review → authorization → CMMS transfer. Each stage needs a stop rule. An ambiguous asset, obsolete document, missing safety plan, or unresolved unit should remain “unconfirmed.” The output schema should have separate fields for asset ID, objective, task array, parts, tools, craft, safety-plan reference, source array, and unresolved questions. This discourages the model from filling every blank with a guess.

A maintenance planner reviews the asset and scope, a safety lead reviews hazards and permits, a stores or purchasing lead checks parts, and a site lead checks the maintenance window and personnel. Roles may be combined in a small plant, but the record should identify who approved which decision. Only reviewed data should be written to the CMMS. IBM’s work order statuses distinguish WAPPR, waiting for approval, from APPR, approved so work can begin. AI output should never be treated as an automatic APPR decision.

Start integration with reviewed import or manual entry to expose mapping and permission problems. When moving to an API, use the draft ID as an idempotency key, reconcile ambiguous API responses against the CMMS, and retain the approved version and transferred payload for later comparison.

Keep safety citation separate from safety authorization

Do not prompt AI to “write a safe shutdown procedure” for an unfamiliar machine. Isolation points, stored pressure, gravity movement, and connections to adjacent equipment require equipment-specific verification. OSHA’s US rule gives examples of procedural elements such as shutdown, isolation, device application and removal, and verification. It does not establish Thai legal requirements. A work order should link an approved asset-specific safety-plan ID and permit, record the on-site check, and stop when a required plan is missing. AI may detect missing or mismatched references. It may not invent isolation steps, verify that isolation was performed, or approve a permit. Site conditions changing after approval require reassessment before work.

Begin with one asset family

Choose a group with stable asset IDs, approved job plans, maintained documents, and reviewers who can check actual tickets. Leave urgent recovery, recently modified machines, and high-risk jobs without established procedures outside the initial draft scope. Inventory duplicate IDs, aliases, revisions, missing part codes, and safety-plan links before running a pilot. Compare historical draft cases with approved records, while allowing reviewers to correct errors in old records too.

Measure more than writing time: correct asset match, reachable source links, current revision use, part-code match, missing-safety-plan detection, human edits, time to approval, and returned tickets. Investigate each material safety or quality error individually. Retest when the model, retrieval index, asset master, or job-plan content changes.

Estimate value with your own baseline

There is no universal, verified percentage saving for all plants. For an illustrative internal scenario, assume 120 applicable work orders per month, 35 minutes each for manual search and transcription, and 20 minutes each for AI draft review. The arithmetic is 120 × (35 − 20) ÷ 60 = 30 hours per month. These are assumptions, not measured results. Data cleanup, reviewer time, API maintenance, and rework reduce the net benefit. Freed time is not automatically a payroll saving; assess whether it can be used for failure analysis, preventive work, or reduced overtime. Do not expand scope if safety or source quality is below the acceptance threshold.

Connect with existing inspection operations

A missed inspection alert indicates that a scheduled check was not performed. The alert alone does not specify which asset needs repair, what resources to prepare, or who may authorize the job. Conversely, a well-drafted maintenance order cannot detect an inspection that never happened. Connect the two by carrying the asset ID and planned inspection ID into a draft, checking for an existing work order, and asking a planner to confirm the scope. The missed-inspection alert guide covers that upstream control. Keep the missed event, a suspected fault, and the eventual decision in separate fields; their coexistence does not establish causation.

Distinguish maintenance work orders from production instructions

A production instruction system distributes approved assembly procedures and keeps obsolete revisions off the shop floor. Production order change control manages quantities, dates, routing, and sequencing. A maintenance work order describes an individual repair, inspection, or preventive task on a specific asset, with resources, safety review, and execution records. The systems can exchange a shutdown window or a fault signal, but production scheduling consent is not energy-isolation authorization.

Example: drafting a pump bearing work order

Suppose a request says “pump three vibrates; consider bearing replacement.” If “pump three” maps to two assets, the system stops and asks a planner to inspect location and nameplate. Once P-204 is confirmed, it retrieves current documents and approved plans for that asset. A plan for bearing inspection is not silently treated as a bearing replacement plan. The draft shows that task mismatch and asks for another approved plan. It proposes a catalogued item code, tool list, and craft requirements only with supporting references. An absent isolation plan remains a blocker; generic lockout text does not cure it. The request ID, draft ID, human approvals, and resulting CMMS ID remain linked. This example and its IDs are fictional, not field instructions.

The site review can be more explicit. A planner compares P-204’s nameplate, location, and modification history with the asset register and records the match. The observed vibration is kept separate from a diagnosis of bearing failure. If two job plans appear, the planner compares their approved revisions, applicable models, and task scope, then records why one was selected or rejected. Stores checks the item code and actual stock; the maintenance lead checks tools and craft. An AI-estimated quantity remains pending until verified and must not become a purchasing commitment by default.

The safety lead opens the approved asset-specific plan and decides whether it fits current site conditions. A plan may exist yet need review after equipment modification or nearby work changes. Authorized personnel separately verify permits and the actual isolation; AI text is not proof that isolation happened. The planner then consolidates the checks, and the approver reviews sources, unresolved fields, the shutdown window, and assigned people. Only after approval does the ticket enter the CMMS, linked to the originating request. The pilot should verify that any missing mandatory check keeps the ticket waiting for approval.

At the next shift handover, the outgoing and incoming teams should use the same ticket to review open items, part receipt, and the available shutdown window. If site conditions changed, they should record the difference and seek any required renewed approval before resuming work. After completion, actual parts, labor, and discovered faults go back into the CMMS as evidence for future job-plan review.

Acceptance tests from pilot to operation

A request for proposal should include realistic test cases and acceptance criteria, rather than asking vendors only for a persuasive demonstration. Include a normal case with one asset ID, but also an alias matching two machines, an obsolete drawing, a substitute part without approval, and a missing safety plan. For each case ask the supplier to show the input, source document, output, stop condition, and reproducible log. A fluent draft is not enough to prove that the system handles uncertainty safely.

In the predeployment factory acceptance test (FAT), use a controlled environment resembling production data. Check asset matching, document revision selection, differences from an approved job plan, parts, tools, craft, approval permissions, and duplicate prevention when writing to the CMMS. Include negative tests: do not finalize an ambiguous asset; warn or hold when no current document exists; do not release work with an unapproved safety plan; and do not move a draft to a work-ready CMMS status before approval. The plant should define its pass thresholds in writing before testing, based on its own risk assessment. These are design examples, not universal regulatory requirements.

At site acceptance testing (SAT), use the actual terminals, network, display languages, access rights, machine nameplates, and shutdown-request process. If Japanese and Thai nicknames refer to one machine, test that both resolve to the same asset ID. On a network interruption, an unsent draft must not appear sent. A night-shift user without approval rights must not issue a work-ready ticket. If a published work order changes, verify revision and renewed approval history. Passing FAT does not guarantee the same behavior under site conditions.

For every test, retain the request, confirmed asset ID, source ID and revision, job plan, AI draft, human edits, approver, and final CMMS record. Set access and retention under company information policies because these records may contain personal or equipment-sensitive data. The point is to be able to reconstruct why a work order was approved and investigate an error; storing every model response indefinitely is not automatically necessary.

Train users and prepare exception routes

Train requesters to record asset number and observed symptom separately. Train planners to check revisions, asset IDs, and part codes. Train approvers to open evidence and recognize a hold condition. Train IT staff to reconcile failed transfers before resending. Early in deployment, record why a proposal was rejected as well as why it was accepted: rejection reasons often expose missing masters or obsolete documents. Maintain a glossary for each site language and keep job-plan names and asset codes stable across translation.

Define a fallback when a newly installed machine lacks an asset record, a vendor manual has no known revision, an emergency shutdown occurs, connectivity fails, the CMMS is unavailable, or a high-risk permit condition changes. Route such work through the site’s established human or paper process, then reconcile the record later. If exceptions cluster on one asset family, remove it from the AI scope until the master data is repaired. Count exceptions by cause; this helps prioritize the next improvement.

Changes after launch require control. A job-plan revision, asset modification, approved part substitute, search-index rebuild, or model change should trigger a relevant regression test. After a document revision, check that obsolete sections are not preferred and identify unreleased work orders potentially affected. Show them to a responsible person for review rather than silently rewriting approved instructions. Classify reported errors as master-data, document, authorization, interface, or model issues before deciding what to fix.

Procurement checklist

Ask vendors to demonstrate asset-ID matching, section-level citations and revisions, approved-plan comparison, item-code handling, stop rules for missing safety plans, role-based approval, reviewed CMMS write-back, and retesting after changes. In a multilingual plant, translated work instructions must preserve equipment names, part codes, units, and safety meaning, with human review. Do not infer local permits or qualifications from general model knowledge.

Frequently asked questions

Q1. Can a plant start without a CMMS?

Yes, if it defines a controlled asset register, document versions, approval owner, and change history. Otherwise free-text tickets become difficult to trace later.

Q2. Can AI write lockout or energy-isolation steps?

This design excludes that function. AI may reference an approved, applicable procedure; authorized site personnel determine and verify the actual steps. The OSHA source is a US example.

Q3. Is a job plan the same as a work order?

No. The job plan is reusable standard content; an individual work order has its own work plan and execution record. IBM documents how applying a job plan creates a work plan that can then be edited without changing the original.

Q4. What differs from equipment-manual search AI?

Search finds candidate sections. Work-order drafting connects verified sections to asset ID, resources, safety plan, approval, and CMMS fields. A search result is not automatically an instruction to perform work.

Q5. How should success be judged?

Compare similar work orders using time, asset match, valid source revision, part match, safety gap detection, review edits, returns, and approval time. Use your plant’s baseline.

Take the next step

Useful maintenance work-order AI produces a traceable proposal for responsible people, not an automatic authorization. TOMAS TECH can review your asset register, document versions, CMMS fields, and approval flow, then select a limited pilot with measurable acceptance criteria. Bring one representative work order to a consultation to identify what may be drafted and what must stay with site personnel.

Sources

• IBM Job Plans
• IBM Creating work orders
• IBM Work order statuses
• IBM Job plans and work orders
• US OSHA 1910.147 (US regulation; not Thai law)
• NIST AI 600-1